Computation and Language (cs.CL)

Abstract: Sentence-by-sentence information extraction from long documents is an exhausting and error-prone task. As the indicator of document skeleton, catalogs naturally chunk documents into segments and provide informative cascade semantics, which can help to reduce the search space. Despite their usefulness, catalogs are hard to be extracted without the assist from external knowledge. For documents that adhere to a specific template, regular expressions are practical to extract catalogs. However, handcrafted heuristics are not applicable when processing documents from different sources with diverse formats. To address this problem, we build a large manually annotated corpus, which is the first dataset for the Catalog Extraction from Documents (CED) task. Based on this corpus, we propose a transition-based framework for parsing documents into catalog trees. The experimental results demonstrate that our proposed method outperforms baseline systems and shows a good ability to transfer. We believe the CED task could fill the gap between raw text segments and information extraction tasks on extremely long documents. Data and code are available at \url{https://github.com/Spico197/CatalogExtraction}

Abstract: With the wide application of Large Language Models (LLMs) such as ChatGPT, how to make the contents generated by LLM accurate and credible becomes very important, especially in complex knowledge-intensive tasks. In this paper, we propose a novel framework called Search-in-the-Chain (SearChain) to improve the accuracy, credibility and traceability of LLM-generated content for multi-hop question answering, which is a typical complex knowledge-intensive task. SearChain is a framework that deeply integrates LLM and information retrieval (IR). In SearChain, LLM constructs a chain-of-query, which is the decomposition of the multi-hop question. Each node of the chain is a query-answer pair consisting of an IR-oriented query and the answer generated by LLM for this query. IR verifies, completes, and traces the information of each node of the chain, so as to guide LLM to construct the correct chain-of-query, and finally answer the multi-hop question. SearChain makes LLM change from trying to give a answer to trying to construct the chain-of-query when faced with the multi-hop question, which can stimulate the knowledge-reasoning ability and provides the interface for IR to be deeply involved in reasoning process of LLM. IR interacts with each node of chain-of-query of LLM. It verifies the information of the node and provides the unknown knowledge to LLM, which ensures the accuracy of the whole chain in the process of LLM generating the answer. Besides, the contents returned by LLM to the user include not only the final answer but also the reasoning process for the question, that is, the chain-of-query and the supporting documents retrieved by IR for each node of the chain, which improves the credibility and traceability of the contents generated by LLM. Experimental results show SearChain outperforms related baselines on four multi-hop question-answering datasets.

Abstract: We propose a novel approach to learn domain-specific plausible materials for components in the vehicle repair domain by probing Pretrained Language Models (PLMs) in a cloze task style setting to overcome the lack of annotated datasets. We devise a new method to aggregate salient predictions from a set of cloze query templates and show that domain-adaptation using either a small, high-quality or a customized Wikipedia corpus boosts performance. When exploring resource-lean alternatives, we find a distilled PLM clearly outperforming a classic pattern-based algorithm. Further, given that 98% of our domain-specific components are multiword expressions, we successfully exploit the compositionality assumption as a way to address data sparsity.

Abstract: Transformer-based language models (LMs) are known to capture factual knowledge in their parameters. While previous work looked into where factual associations are stored, only little is known about how they are retrieved internally during inference. We investigate this question through the lens of information flow. Given a subject-relation query, we study how the model aggregates information about the subject and relation to predict the correct attribute. With interventions on attention edges, we first identify two critical points where information propagates to the prediction: one from the relation positions followed by another from the subject positions. Next, by analyzing the information at these points, we unveil a three-step internal mechanism for attribute extraction. First, the representation at the last-subject position goes through an enrichment process, driven by the early MLP sublayers, to encode many subject-related attributes. Second, information from the relation propagates to the prediction. Third, the prediction representation "queries" the enriched subject to extract the attribute. Perhaps surprisingly, this extraction is typically done via attention heads, which often encode subject-attribute mappings in their parameters. Overall, our findings introduce a comprehensive view of how factual associations are stored and extracted internally in LMs, facilitating future research on knowledge localization and editing.

Abstract: Universal Information Extraction (UIE) is an area of interest due to the challenges posed by varying targets, heterogeneous structures, and demand-specific schemas. However, previous works have only achieved limited success by unifying a few tasks, such as Named Entity Recognition (NER) and Relation Extraction (RE), which fall short of being authentic UIE models particularly when extracting other general schemas such as quadruples and quintuples. Additionally, these models used an implicit structural schema instructor, which could lead to incorrect links between types, hindering the model's generalization and performance in low-resource scenarios. In this paper, we redefine the authentic UIE with a formal formulation that encompasses almost all extraction schemas. To the best of our knowledge, we are the first to introduce UIE for any kind of schemas. In addition, we propose RexUIE, which is a Recursive Method with Explicit Schema Instructor for UIE. To avoid interference between different types, we reset the position ids and attention mask matrices. RexUIE shows strong performance under both full-shot and few-shot settings and achieves State-of-the-Art results on the tasks of extracting complex schemas.

Abstract: This paper provides a detailed discussion of the multilingual tokenizer used for GPT-SW3. It was trained on the Nordic Pile using the SentencePiece library and the BPE algorithm. We outline the tokenizer's most important features and share details on its learned vocabulary. In addition, we systematically analyze the properties and evaluate the performance of the tokenizer with regard to the different languages present in the data.

Abstract: Dense vector representations for textual data are crucial in modern NLP. Word embeddings and sentence embeddings estimated from raw texts are key in achieving state-of-the-art results in various tasks requiring semantic understanding. However, obtaining embeddings at the document level is challenging due to computational requirements and lack of appropriate data. Instead, most approaches fall back on computing document embeddings based on sentence representations. Although there exist architectures and models to encode documents fully, they are in general limited to English and few other high-resourced languages. In this work, we provide a systematic comparison of methods to produce document-level representations from sentences based on LASER, LaBSE, and Sentence BERT pre-trained multilingual models. We compare input token number truncation, sentence averaging as well as some simple windowing and in some cases new augmented and learnable approaches, on 3 multi- and cross-lingual tasks in 8 languages belonging to 3 different language families. Our task-based extrinsic evaluations show that, independently of the language, a clever combination of sentence embeddings is usually better than encoding the full document as a single unit, even when this is possible. We demonstrate that while a simple sentence average results in a strong baseline for classification tasks, more complex combinations are necessary for semantic tasks.

Abstract: Transformer networks have become the preferred architecture for many tasks due to their state-of-the-art performance. However, the optimal way to implement residual connections in Transformer, which are essential for effective training, is still debated. Two widely used variants are the Post-Layer-Normalization (Post-LN) and Pre-Layer-Normalization (Pre-LN) Transformers, which apply layer normalization after each residual block's output or before each residual block's input, respectively. While both variants enjoy their advantages, they also suffer from severe limitations: Post-LN causes gradient vanishing issue that hinders training deep Transformers, and Pre-LN causes representation collapse issue that limits model capacity. In this paper, we propose ResiDual, a novel Transformer architecture with Pre-Post-LN (PPLN), which fuses the connections in Post-LN and Pre-LN together and inherits their advantages while avoids their limitations. We conduct both theoretical analyses and empirical experiments to verify the effectiveness of ResiDual. Theoretically, we prove that ResiDual has a lower bound on the gradient to avoid the vanishing issue due to the residual connection from Pre-LN. Moreover, ResiDual also has diverse model representations to avoid the collapse issue due to the residual connection from Post-LN. Empirically, ResiDual outperforms both Post-LN and Pre-LN on several machine translation benchmarks across different network depths and data sizes. Thanks to the good theoretical and empirical performance, ResiDual Transformer can serve as a foundation architecture for different AI models (e.g., large language models). Our code is available at https://github.com/microsoft/ResiDual.

Abstract: NLP datasets annotated with human judgments are rife with disagreements between the judges. This is especially true for tasks depending on subjective judgments such as sentiment analysis or offensive language detection. Particularly in these latter cases, the NLP community has come to realize that the approach of 'reconciling' these different subjective interpretations is inappropriate. Many NLP researchers have therefore concluded that rather than eliminating disagreements from annotated corpora, we should preserve them-indeed, some argue that corpora should aim to preserve all annotator judgments. But this approach to corpus creation for NLP has not yet been widely accepted. The objective of the LeWiDi series of shared tasks is to promote this approach to developing NLP models by providing a unified framework for training and evaluating with such datasets. We report on the second LeWiDi shared task, which differs from the first edition in three crucial respects: (i) it focuses entirely on NLP, instead of both NLP and computer vision tasks in its first edition; (ii) it focuses on subjective tasks, instead of covering different types of disagreements-as training with aggregated labels for subjective NLP tasks is a particularly obvious misrepresentation of the data; and (iii) for the evaluation, we concentrate on soft approaches to evaluation. This second edition of LeWiDi attracted a wide array of participants resulting in 13 shared task submission papers.

Abstract: This paper aims to quantitatively evaluate the performance of ChatGPT, an interactive large language model, on inter-sentential relations such as temporal relations, causal relations, and discourse relations. Given ChatGPT's promising performance across various tasks, we conduct extensive evaluations on the whole test sets of 13 datasets, including temporal and causal relations, PDTB2.0-based and dialogue-based discourse relations, and downstream applications on discourse understanding. To achieve reliable results, we adopt three tailored prompt templates for each task, including the zero-shot prompt template, zero-shot prompt engineering (PE) template, and in-context learning (ICL) prompt template, to establish the initial baseline scores for all popular sentence-pair relation classification tasks for the first time. We find that ChatGPT exhibits strong performance in detecting and reasoning about causal relations, while it may not be proficient in identifying the temporal order between two events. It can recognize most discourse relations with existing explicit discourse connectives, but the implicit discourse relation still remains a challenging task. Meanwhile, ChatGPT performs poorly in the dialogue discourse parsing task that requires structural understanding in a dialogue before being aware of the discourse relation.

Abstract: Online propaganda poses a severe threat to the integrity of societies. However, existing datasets for detecting online propaganda have a key limitation: they were annotated using weak labels that can be noisy and even incorrect. To address this limitation, our work makes the following contributions: (1) We present \dataset: a novel dataset (N=30,000) for detecting online propaganda with high-quality labels. To the best of our knowledge, \dataset is the first dataset for detecting online propaganda that was created through human annotation. (2) We show empirically that state-of-the-art language models fail in detecting online propaganda when trained with weak labels (AUC: 64.03). In contrast, state-of-the-art language models can accurately detect online propaganda when trained with our high-quality labels (AUC: 92.25), which is an improvement of ~44%. (3) To address the cost of labeling, we extend our work to few-shot learning. Specifically, we show that prompt-based learning using a small sample of high-quality labels can still achieve a reasonable performance (AUC: 80.27). Finally, we discuss implications for the NLP community to balance the cost and quality of labeling. Crucially, our work highlights the importance of high-quality labels for sensitive NLP tasks such as propaganda detection.

Abstract: Advances in the Visually-rich Document Understanding (VrDU) field and particularly the Key-Information Extraction (KIE) task are marked with the emergence of efficient Transformer-based approaches such as the LayoutLM models. Despite the good performance of KIE models when fine-tuned on public benchmarks, they still struggle to generalize on complex real-life use-cases lacking sufficient document annotations. Our research highlighted that KIE standard benchmarks such as SROIE and FUNSD contain significant similarity between training and testing documents and can be adjusted to better evaluate the generalization of models. In this work, we designed experiments to quantify the information redundancy in public benchmarks, revealing a 75% template replication in SROIE official test set and 16% in FUNSD. We also proposed resampling strategies to provide benchmarks more representative of the generalization ability of models. We showed that models not suited for document analysis struggle on the adjusted splits dropping on average 10,5% F1 score on SROIE and 3.5% on FUNSD compared to multi-modal models dropping only 7,5% F1 on SROIE and 0.5% F1 on FUNSD.

Abstract: In recent years, the field of document understanding has progressed a lot. A significant part of this progress has been possible thanks to the use of language models pretrained on large amounts of documents. However, pretraining corpora used in the domain of document understanding are single domain, monolingual, or nonpublic. Our goal in this paper is to propose an efficient pipeline for creating a big-scale, diverse, multilingual corpus of PDF files from all over the Internet using Common Crawl, as PDF files are the most canonical types of documents as considered in document understanding. We analysed extensively all of the steps of the pipeline and proposed a solution which is a trade-off between data quality and processing time. We also share a CCpdf corpus in a form or an index of PDF files along with a script for downloading them, which produces a collection useful for language model pretraining. The dataset and tools published with this paper offer researchers the opportunity to develop even better multilingual language models.

Abstract: As foundation models continue to exponentially scale in size, efficient methods of adaptation become increasingly critical. Parameter-efficient fine-tuning (PEFT), a recent class of techniques that require only modifying a small percentage of the model parameters, is currently the most popular method for adapting large language models (LLMs). Several PEFT techniques have recently been proposed with varying tradeoffs. We provide a comprehensive and uniform benchmark of various PEFT techniques across a representative LLM, the FLAN-T5 model, and evaluate model performance across different data scales of classification and generation datasets. Based on this, we provide a framework for choosing the optimal fine-tuning techniques given the task type and data availability. Contrary to popular belief, we also empirically prove that PEFT techniques converge slower than full tuning in low data scenarios, and posit the amount of data required for PEFT methods to both perform well and converge efficiently. Lastly, we further optimize these PEFT techniques by selectively choosing which parts of the model to train, and find that these techniques can be applied with significantly fewer parameters while maintaining and even improving performance.